Re: what's the proper value for BB_NUMBER_THREADS?

[Darren Hart <dvhart@linux.intel.com>]

On 10/31/2011 10:39 AM, Gary Thomas wrote:
> On 2011-10-31 11:25, Robert P. J. Day wrote:
>> On Mon, 31 Oct 2011, Mark Hatle wrote:
>>
>>> On 10/30/11 11:15 AM, Robert P. J. Day wrote:
>>>> On Sun, 30 Oct 2011, Christian Gagneraud wrote:
>>>>
>>>>> On 30/10/11 15:32, Robert P. J. Day wrote:
>>>>>>
>>>>>>     all the docs recommend twice the number of cores (AFAICT), yet the
>>>>>> template local.conf file suggests that, for a quad core, the value of
>>>>>> 4 would be appropriate.  shouldn't that say 8?  same with
>>>>>> PARALLEL_MAKE?
>>>>>
>>>>> Hi Robert,
>>>>>
>>>>> The Poky ref manual says (rule of thumb) x2 for BB_NUMBER_THREADS,
>>>>> and x1.5 for PARALLEL_MAKE.
>>>>
>>>>    if that's the case, anyone object to my submitting a patch to
>>>> update local.conf.sample appropriately?
>>>>
>>>> rday
>>>>
>>>
>>> I agree the manual and local.conf files should match.  The issue
>>> seems to be that the perfect values are subjective.  Things like
>>> memory, disk speed, chipset latency, and of course processor
>>> speed/cores all affect the optimal setting. But we do need a
>>> consistent rule of thumb..  I myself usually use x2 for both THREADS
>>> and MAKE.
>>
>>    that's what i would normally use, assuming that having an overly
>> high value for either of those settings can't possibly hurt.  if
>> that's the consensus, i can adjust the references to say 2x everywhere
>> that i know of.  just let me know.
> 
> Look back in the archives - I think it was Richard that did a fairly
> extensive study of this and he (whoever it was) found that there were
> saturation points and trying to get beyond them had a negative impact.

2x on each works well up to about 12 in my experience. Richard has found
benefits using 24 on a similar system with more memory. However, if you
aren't building on a monster machine, then 2x serves as a reasonable
rule of thumb.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel